Re: [PATCH v2 1/1] memory tier: consolidate the initialization of memory tiers

[Ho-Ren (Jack) Chuang]

Hi Jonathan,

I appreciate your feedback and valuable suggestions. Replies inlined.


July 2, 2024 at 6:25 AM, "Jonathan Cameron" <Jonathan.Cameron@huawei.com> wrote:
> 
> On Fri, 28 Jun 2024 06:09:23 +0000
> 
> "Ho-Ren (Jack) Chuang" <horen.chuang@linux.dev> wrote:
> 
> > 
> > If we simply move the set_node_memory_tier() from memory_tier_init()
> > 
> >  to late_initcall(), it will result in HMAT not registering
> > 
> >  the mt_adistance_algorithm callback function, because
> > 
> >  set_node_memory_tier() is not performed during the memory tiering
> > 
> >  initialization phase, leading to a lack of correct default_dram
> > 
> >  information.
> > 
> >  
> > 
> >  Therefore, we introduced a nodemask to pass the information of the
> > 
> >  default DRAM nodes. The reason for not choosing to reuse
> > 
> >  default_dram_type->nodes is that it is not clean enough. So in the end,
> > 
> >  we use a __initdata variable, which is a variable that is released once
> > 
> >  initialization is complete, including both CPU and memory nodes for HMAT
> > 
> >  to iterate through.
> > 
> >  
> > 
> >  Besides, since default_dram_type may be checked/used during the
> > 
> >  initialization process of HMAT and drivers, it is better to keep the
> > 
> >  allocation of default_dram_type in memory_tier_init().
> > 
> >  
> > 
> >  Signed-off-by: Ho-Ren (Jack) Chuang <horenchuang@bytedance.com>
> > 
> >  Suggested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
> > 
> >  ---
> > 
> >  drivers/acpi/numa/hmat.c | 5 +--
> > 
> >  include/linux/memory-tiers.h | 2 ++
> > 
> >  mm/memory-tiers.c | 59 +++++++++++++++---------------------
> > 
> >  3 files changed, 28 insertions(+), 38 deletions(-)
> > 
> >  
> > 
> >  diff --git a/drivers/acpi/numa/hmat.c b/drivers/acpi/numa/hmat.c
> > 
> >  index 2c8ccc91ebe6..a2f9e7a4b479 100644
> > 
> >  --- a/drivers/acpi/numa/hmat.c
> > 
> >  +++ b/drivers/acpi/numa/hmat.c
> > 
> >  @@ -940,10 +940,7 @@ static int hmat_set_default_dram_perf(void)
> > 
> >  struct memory_target *target;
> > 
> >  struct access_coordinate *attrs;
> > 
> >  
> > 
> >  - if (!default_dram_type)
> > 
> >  - return -EIO;
> > 
> >  -
> > 
> >  - for_each_node_mask(nid, default_dram_type->nodes) {
> > 
> >  + for_each_node_mask(nid, default_dram_nodes) {
> > 
> 
> As below. Do we care if the combination of RAM + CPU wasn't true
> 
> earlier and is true by this point? If not, why not just
> 
> compute the node mask in here and not store it.
> 

It makes sense to me. I think we can move the computation to here
and remove the global node mask.

> > 
> > pxm = node_to_pxm(nid);
> > 
> >  target = find_mem_target(pxm);
> > 
> >  if (!target)
> > 
> >  diff --git a/include/linux/memory-tiers.h b/include/linux/memory-tiers.h
> > 
> >  index 0d70788558f4..fa61ad9c4d75 100644
> > 
> >  --- a/include/linux/memory-tiers.h
> > 
> >  +++ b/include/linux/memory-tiers.h
> > 
> >  @@ -38,6 +38,7 @@ struct access_coordinate;
> > 
> >  #ifdef CONFIG_NUMA
> > 
> >  extern bool numa_demotion_enabled;
> > 
> >  extern struct memory_dev_type *default_dram_type;
> > 
> >  +extern nodemask_t default_dram_nodes __initdata;
> > 
> >  struct memory_dev_type *alloc_memory_type(int adistance);
> > 
> >  void put_memory_type(struct memory_dev_type *memtype);
> > 
> >  void init_node_memory_type(int node, struct memory_dev_type *default_type);
> > 
> >  @@ -76,6 +77,7 @@ static inline bool node_is_toptier(int node)
> > 
> >  
> > 
> >  #define numa_demotion_enabled false
> > 
> >  #define default_dram_type NULL
> > 
> >  +#define default_dram_nodes NODE_MASK_NONE
> > 
> >  /*
> > 
> >  * CONFIG_NUMA implementation returns non NULL error.
> > 
> >  */
> > 
> >  diff --git a/mm/memory-tiers.c b/mm/memory-tiers.c
> > 
> >  index 6632102bd5c9..a19a90c3ad36 100644
> > 
> >  --- a/mm/memory-tiers.c
> > 
> >  +++ b/mm/memory-tiers.c
> > 
> >  @@ -43,6 +43,7 @@ static LIST_HEAD(memory_tiers);
> > 
> >  static LIST_HEAD(default_memory_types);
> > 
> >  static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
> > 
> >  struct memory_dev_type *default_dram_type;
> > 
> >  +nodemask_t default_dram_nodes __initdata = NODE_MASK_NONE;
> > 
> >  
> > 
> >  static const struct bus_type memory_tier_subsys = {
> > 
> >  .name = "memory_tiering",
> > 
> >  @@ -671,28 +672,38 @@ EXPORT_SYMBOL_GPL(mt_put_memory_types);
> > 
> >  
> > 
> >  /*
> > 
> >  * This is invoked via `late_initcall()` to initialize memory tiers for
> > 
> >  - * CPU-less memory nodes after driver initialization, which is
> > 
> >  - * expected to provide `adistance` algorithms.
> > 
> >  + * memory nodes, both with and without CPUs. After the initialization of
> > 
> >  + * firmware and devices, adistance algorithms are expected to be provided.
> > 
> >  */
> > 
> >  static int __init memory_tier_late_init(void)
> > 
> >  {
> > 
> >  int nid;
> > 
> >  + struct memory_tier *memtier;
> > 
> >  
> > 
> >  + get_online_mems();
> > 
> >  guard(mutex)(&memory_tier_lock);
> > 
> >  + /*
> > 
> >  + * Look at all the existing and uninitialized N_MEMORY nodes and
> > 
> >  + * add them to default memory tier or to a tier if we already have
> > 
> >  + * memory types assigned.
> > 
> >  + */
> > 
> >  for_each_node_state(nid, N_MEMORY) {
> > 
> >  /*
> > 
> >  - * Some device drivers may have initialized memory tiers
> > 
> >  - * between `memory_tier_init()` and `memory_tier_late_init()`,
> > 
> >  - * potentially bringing online memory nodes and
> > 
> >  - * configuring memory tiers. Exclude them here.
> > 
> >  + * Some device drivers may have initialized
> > 
> >  + * memory tiers, potentially bringing memory nodes
> > 
> >  + * online and configuring memory tiers.
> > 
> >  + * Exclude them here.
> > 
> >  */
> > 
> >  if (node_memory_types[nid].memtype)
> > 
> >  continue;
> > 
> >  
> > 
> >  - set_node_memory_tier(nid);
> > 
> >  + memtier = set_node_memory_tier(nid);
> > 
> >  + if (IS_ERR(memtier))
> > 
> >  + /* Continue with memtiers we are able to setup. */
> > 
> 
> Might later ones be possible if we just continued this loop?
> 

I agree with you that theoretically, it’s possible
for later attempts to succeed. I also agree that
there is no harm in iterating through all possibilities.
Therefore, we can do a continue here.

Since it's legacy code.
I would also like to hear Huang, Ying’s thoughts about this.

> > 
> > + break;
> > 
> >  }
> > 
> >  -
> > 
> 
> White space was harmless - I'd leave it there rather than adding noise to 
> this diff.
> 

Thanks! Got it. I will roll it back in the v3.

> > 
> > establish_demotion_targets();
> > 
> >  + put_online_mems();
> > 
> >  
> > 
> >  return 0;
> > 
> >  }
> > 
> >  @@ -875,8 +886,7 @@ static int __meminit memtier_hotplug_callback(struct 
> > notifier_block *self,
> > 
> >  
> > 
> >  static int __init memory_tier_init(void)
> > 
> >  {
> > 
> >  - int ret, node;
> > 
> >  - struct memory_tier *memtier;
> > 
> >  + int ret;
> > 
> >  
> > 
> >  ret = subsys_virtual_register(&memory_tier_subsys, NULL);
> > 
> >  if (ret)
> > 
> >  @@ -887,7 +897,8 @@ static int __init memory_tier_init(void)
> > 
> >  GFP_KERNEL);
> > 
> >  WARN_ON(!node_demotion);
> > 
> >  #endif
> > 
> >  - mutex_lock(&memory_tier_lock);
> > 
> >  +
> > 
> >  + guard(mutex)(&memory_tier_lock);
> > 
> 
> If this was safe to do without the rest of the change (I think so)
> 
> then better to pull that out as a trivial precursor so less noise
> 
> in here.
> 

Do you mean instead of using guard(mutex)(),
use mutex_lock() as it was? or?

> > 
> > /*
> > 
> >  * For now we can have 4 faster memory tiers with smaller adistance
> > 
> >  * than default DRAM tier.
> > 
> >  @@ -897,29 +908,9 @@ static int __init memory_tier_init(void)
> > 
> >  if (IS_ERR(default_dram_type))
> > 
> >  panic("%s() failed to allocate default DRAM tier\n", __func__);
> > 
> >  
> > 
> >  - /*
> > 
> >  - * Look at all the existing N_MEMORY nodes and add them to
> > 
> >  - * default memory tier or to a tier if we already have memory
> > 
> >  - * types assigned.
> > 
> >  - */
> > 
> >  - for_each_node_state(node, N_MEMORY) {
> > 
> >  - if (!node_state(node, N_CPU))
> > 
> >  - /*
> > 
> >  - * Defer memory tier initialization on
> > 
> >  - * CPUless numa nodes. These will be initialized
> > 
> >  - * after firmware and devices are initialized.
> > 
> >  - */
> > 
> >  - continue;
> > 
> >  -
> > 
> >  - memtier = set_node_memory_tier(node);
> > 
> >  - if (IS_ERR(memtier))
> > 
> >  - /*
> > 
> >  - * Continue with memtiers we are able to setup
> > 
> >  - */
> > 
> >  - break;
> > 
> >  - }
> > 
> >  - establish_demotion_targets();
> > 
> >  - mutex_unlock(&memory_tier_lock);
> > 
> >  + /* Record nodes with memory and CPU to set default DRAM performance. */
> > 
> >  + nodes_and(default_dram_nodes, node_states[N_MEMORY],
> > 
> >  + node_states[N_CPU]);
> > 
> 
> There are systems where (for various esoteric reasons, such as describing an
> 
> association with some other memory that isn't DRAM where the granularity
> 
> doesn't match) the CPU nodes contain no DRAM but rather it's one node away.
> 
> Handling that can be a job for another day though.
> 

Thank you for informing me of this situation.
Sounds like handling that also requires a mapping table between
the CPU and the corresponding DRAM.

> Why does this need to be computed here? Why not do it in
> 
> hmat_set_default_dram_perf? Doesn't seem to be used anywhere else.
> 
Replied above.
> > 
> > hotplug_memory_notifier(memtier_hotplug_callback, MEMTIER_HOTPLUG_PRI);
> > 
> >  return 0;
> >
>

--
Best Regards,
Ho-Ren (Jack) Chuang